Data augmentation with GAN: Improving Chest X-Ray pathologies prediction on class-imbalanced cases

Abstract

When one applies machine learning to a real-world problem, sometimes data imbalance makes a crucial impact on the resulting model’s performance. We propose to use generative adversarial network (GAN) to do data balancing through data augmentation in data preprocessing step of binary classification task. We train CycleGAN on unpaired images to be able to produce images from the opposite class for any given input image. After training we use it to produce images from the opposite class for every image in a given imbalanced dataset, thus making it fully-balanced. The proposed augmentation technique may be used as the preprocessing step in binary classification tasks. We show that it improves performance in pneumonia presence/absence classification task on X-ray images. We inspect how binary classifier performance changes if the dataset used for GAN training differs from the dataset we measure binary classifier performance on. We also inspect its behavior on several other binary classification tasks related to medical imaging.